Zaliczenie - ImageNet
Mateusz Pieszczek
15 06 2022
library(keras)
library(jpeg)Import modeli.
model_res <- application_resnet50(weights = 'imagenet') #modele z konkursu
model_vgg16 <- application_vgg16(weights = 'imagenet')
model_vgg19 <- application_vgg19(weights = 'imagenet')Funkcja do plotowania pobranych rysunków.
plot_img <- function(obraz) {
img_path <- paste0("obrazy/", obraz)
plot.new()
rasterImage(readJPEG(img_path), 0, 0, 1, 1)
}Funkcję, która automatycznie rozpoznaje obraz i podaje wyniki uzyskany przez modele.
what_is_it <- function(obraz,n=5) {
img_path <- paste0("obrazy/", obraz)
img <- image_load(img_path, target_size = c(224,224))
x <- image_to_array(img)
x <- array_reshape(x, c(1, dim(x)))
x <- imagenet_preprocess_input(x)
preds <- model_res %>% predict(x)
cat("ResNet50\n\n")
print(imagenet_decode_predictions(preds, top = n)[[1]])
cat("\n\n")
preds <- model_vgg16 %>% predict(x)
cat("VGG16\n\n")
print(imagenet_decode_predictions(preds, top = n)[[1]])
cat("\n\n")
preds <- model_vgg19 %>% predict(x)
cat("VGG19\n\n")
print(imagenet_decode_predictions(preds, top = n)[[1]])
}Użyłem zdjęć które akurat miałem na komputerze. Starałem się wybrać specjalnie takie zdjęcia, które modele może rozpozna. # “Tygrys” - Andy Worhol. Zwierzę nie zostało rozpoznane, najwidoczniej ten obraz jest na swój sposób zbyt abstrakcyjny.
plot_img('2.jpeg')
what_is_it('2.jpeg',10)## ResNet50
##
## class_name class_description score
## 1 n03724870 mask 0.44941467
## 2 n03249569 drum 0.08402407
## 3 n04192698 shield 0.07900365
## 4 n04476259 tray 0.05358854
## 5 n04522168 vase 0.04930428
## 6 n06596364 comic_book 0.02649498
## 7 n03637318 lampshade 0.02105241
## 8 n04525038 velvet 0.01446214
## 9 n03595614 jersey 0.01267335
## 10 n02910353 buckle 0.01250398
##
##
## VGG16
##
## class_name class_description score
## 1 n04192698 shield 0.09989087
## 2 n04522168 vase 0.09825062
## 3 n03063599 coffee_mug 0.04439355
## 4 n03950228 pitcher 0.04378205
## 5 n03724870 mask 0.03829108
## 6 n03595614 jersey 0.03577395
## 7 n03929855 pickelhaube 0.03515862
## 8 n07248320 book_jacket 0.03455539
## 9 n01950731 sea_slug 0.02938972
## 10 n06596364 comic_book 0.02928065
##
##
## VGG19
##
## class_name class_description score
## 1 n04192698 shield 0.08880865
## 2 n04522168 vase 0.05861242
## 3 n02910353 buckle 0.04523376
## 4 n04476259 tray 0.04186505
## 5 n06596364 comic_book 0.03834595
## 6 n03724870 mask 0.03822150
## 7 n07248320 book_jacket 0.03201891
## 8 n02865351 bolo_tie 0.02429623
## 9 n03950228 pitcher 0.02122514
## 10 n03595614 jersey 0.01953672“3 krzyże” - zdjęcie autorstwa Andy’ego Worhola .
Tu tez model daje nam wyniki nie pewne i raczej nie adekwatne. Z pewniejszych mamy gwóźdź, kopertę. Z mających jakichś sens jest ołtarz, ale ma jedynie score = 0.1. taki sam wynik mają światła uliczne.
plot_img('3.jpeg')
what_is_it('3.jpeg',10)## ResNet50
##
## class_name class_description score
## 1 n04548280 wall_clock 0.18670860
## 2 n02699494 altar 0.10069993
## 3 n06874185 traffic_light 0.09749167
## 4 n03998194 prayer_rug 0.09195612
## 5 n02786058 Band_Aid 0.06532117
## 6 n07248320 book_jacket 0.04222833
## 7 n04372370 switch 0.03181378
## 8 n03871628 packet 0.02873185
## 9 n04116512 rubber_eraser 0.02526935
## 10 n03938244 pillow 0.02253191
##
##
## VGG16
##
## class_name class_description score
## 1 n03804744 nail 0.18007764
## 2 n03291819 envelope 0.07592065
## 3 n04548280 wall_clock 0.06813763
## 4 n07248320 book_jacket 0.04944785
## 5 n02708093 analog_clock 0.04319844
## 6 n04118776 rule 0.03151427
## 7 n03595614 jersey 0.02811432
## 8 n03916031 perfume 0.02749477
## 9 n03729826 matchstick 0.02264084
## 10 n03017168 chime 0.02055230
##
##
## VGG19
##
## class_name class_description score
## 1 n03804744 nail 0.35801905
## 2 n03291819 envelope 0.10643652
## 3 n03041632 cleaver 0.03697248
## 4 n03498962 hatchet 0.03248440
## 5 n03481172 hammer 0.03100445
## 6 n07248320 book_jacket 0.02170208
## 7 n04153751 screw 0.02084891
## 8 n04118776 rule 0.02078995
## 9 n04154565 screwdriver 0.01823496
## 10 n03729826 matchstick 0.01589454Zdjęcia czaszek.
Tu spodziewałem się jakiegoś sensownego wyniku. Najwidoczniej nie ma kategorii typu czaszka/kości. Mamy: maskę, triceratopsa, hipopotama, muszlę
plot_img('6.jpg')
what_is_it('6.jpg',10)## ResNet50
##
## class_name class_description score
## 1 n03724870 mask 0.848836124
## 2 n01704323 triceratops 0.109410793
## 3 n01943899 conch 0.010573359
## 4 n03388043 fountain 0.003492411
## 5 n02398521 hippopotamus 0.002495643
## 6 n03584829 iron 0.001536417
## 7 n02877765 bottlecap 0.001477902
## 8 n07768694 pomegranate 0.001471463
## 9 n03041632 cleaver 0.001454158
## 10 n02951585 can_opener 0.001329985
##
##
## VGG16
##
## class_name class_description score
## 1 n01943899 conch 0.31512806
## 2 n02398521 hippopotamus 0.21885042
## 3 n01704323 triceratops 0.10978002
## 4 n03724870 mask 0.04939686
## 5 n02655020 puffer 0.03638939
## 6 n01664065 loggerhead 0.02799696
## 7 n02395406 hog 0.02106269
## 8 n02110341 dalmatian 0.01567828
## 9 n07760859 custard_apple 0.01098730
## 10 n02109047 Great_Dane 0.01020160
##
##
## VGG19
##
## class_name class_description score
## 1 n01943899 conch 0.428744614
## 2 n01704323 triceratops 0.325032413
## 3 n03724870 mask 0.018283293
## 4 n02110341 dalmatian 0.017289015
## 5 n02398521 hippopotamus 0.011839163
## 6 n07860988 dough 0.010742976
## 7 n01494475 hammerhead 0.008715058
## 8 n07760859 custard_apple 0.008697555
## 9 n01491361 tiger_shark 0.008430598
## 10 n02640242 sturgeon 0.008410012plot_img('5.jpeg')
what_is_it('5.jpeg',10)## ResNet50
##
## class_name class_description score
## 1 n01704323 triceratops 0.668920457
## 2 n03724870 mask 0.103897244
## 3 n02395406 hog 0.037166942
## 4 n01871265 tusker 0.020899408
## 5 n03532672 hook 0.018225225
## 6 n02397096 warthog 0.015182097
## 7 n03584829 iron 0.011988674
## 8 n03388043 fountain 0.009407327
## 9 n02396427 wild_boar 0.009094472
## 10 n02138441 meerkat 0.007112049
##
##
## VGG16
##
## class_name class_description score
## 1 n02447366 badger 0.09340165
## 2 n01748264 Indian_cobra 0.09206503
## 3 n13040303 stinkhorn 0.06898924
## 4 n02415577 bighorn 0.06746583
## 5 n02504013 Indian_elephant 0.06126913
## 6 n01704323 triceratops 0.04788818
## 7 n01616318 vulture 0.04015041
## 8 n01742172 boa_constrictor 0.03157759
## 9 n02056570 king_penguin 0.02827452
## 10 n02444819 otter 0.02377487
##
##
## VGG19
##
## class_name class_description score
## 1 n02088238 basset 0.12963785
## 2 n03743016 megalith 0.10288966
## 3 n02100236 German_short-haired_pointer 0.08933358
## 4 n02088632 bluetick 0.07484871
## 5 n01704323 triceratops 0.04566663
## 6 n02504013 Indian_elephant 0.02518542
## 7 n02109047 Great_Dane 0.02389983
## 8 n01871265 tusker 0.02389858
## 9 n03724870 mask 0.02175534
## 10 n02089973 English_foxhound 0.02009206Obraz łódki
Kolejny obraz tym razem pomimo pewnego poziomu abstrakcji udało się modelom rozpoznać łódkę. Choć 2 z modeli klasyfikują ją jako wrak.
plot_img('7.jpg')
what_is_it('7.jpg')## ResNet50
##
## class_name class_description score
## 1 n04606251 wreck 0.50756419
## 2 n03447447 gondola 0.18805397
## 3 n02951358 canoe 0.10990557
## 4 n03873416 paddle 0.05025375
## 5 n01498041 stingray 0.01574951
##
##
## VGG16
##
## class_name class_description score
## 1 n04606251 wreck 0.922342598
## 2 n02951358 canoe 0.021291759
## 3 n03447447 gondola 0.021095062
## 4 n03873416 paddle 0.010799774
## 5 n03947888 pirate 0.003793182
##
##
## VGG19
##
## class_name class_description score
## 1 n03447447 gondola 0.61423272
## 2 n03873416 paddle 0.12657157
## 3 n02951358 canoe 0.12574644
## 4 n04606251 wreck 0.05483563
## 5 n04347754 submarine 0.01953840Zdjęcie kota.
Tutaj jako, że mamy zdjęcie zwierzęcia to poradził sobie dobrze. Poprawnie też zaklasyfikował rasę.
plot_img('8.jpg')
what_is_it('8.jpg')## ResNet50
##
## class_name class_description score
## 1 n02123597 Siamese_cat 0.9944683313
## 2 n02124075 Egyptian_cat 0.0038369833
## 3 n02127052 lynx 0.0003174354
## 4 n04265275 space_heater 0.0001853825
## 5 n03887697 paper_towel 0.0001048127
##
##
## VGG16
##
## class_name class_description score
## 1 n02123597 Siamese_cat 0.957894862
## 2 n02124075 Egyptian_cat 0.013326641
## 3 n02091467 Norwegian_elkhound 0.010017936
## 4 n02127052 lynx 0.004295040
## 5 n02104365 schipperke 0.003103527
##
##
## VGG19
##
## class_name class_description score
## 1 n02123597 Siamese_cat 0.856924534
## 2 n02124075 Egyptian_cat 0.047059707
## 3 n02127052 lynx 0.035420708
## 4 n04265275 space_heater 0.010406610
## 5 n02091467 Norwegian_elkhound 0.007348414Zdjęcie ręki
Jedno zdjęcie mojej ręki i drugie lepiej wyeksponowane. Jak widać oba źle zaklasyfikowane.
plot_img('9.jpg')
what_is_it('9.jpg')## ResNet50
##
## class_name class_description score
## 1 n02786058 Band_Aid 0.10499439
## 2 n03085013 computer_keyboard 0.03806188
## 3 n04372370 switch 0.02732531
## 4 n03804744 nail 0.02462534
## 5 n03937543 pill_bottle 0.02395504
##
##
## VGG16
##
## class_name class_description score
## 1 n02123597 Siamese_cat 0.07616912
## 2 n02443484 black-footed_ferret 0.05281087
## 3 n03188531 diaper 0.03371193
## 4 n02108915 French_bulldog 0.03262693
## 5 n02364673 guinea_pig 0.03035467
##
##
## VGG19
##
## class_name class_description score
## 1 n02786058 Band_Aid 0.19901666
## 2 n03188531 diaper 0.04057379
## 3 n15075141 toilet_tissue 0.03436757
## 4 n03887697 paper_towel 0.03072196
## 5 n03825788 nipple 0.02303490plot_img('hand.jpeg')
what_is_it('hand.jpeg')## ResNet50
##
## class_name class_description score
## 1 n02786058 Band_Aid 0.12205970
## 2 n02951585 can_opener 0.09360551
## 3 n04357314 sunscreen 0.05273188
## 4 n03876231 paintbrush 0.04377431
## 5 n04118776 rule 0.03611134
##
##
## VGG16
##
## class_name class_description score
## 1 n02951585 can_opener 0.16363066
## 2 n02786058 Band_Aid 0.06926748
## 3 n03935335 piggy_bank 0.05141151
## 4 n04357314 sunscreen 0.03837197
## 5 n03876231 paintbrush 0.03771670
##
##
## VGG19
##
## class_name class_description score
## 1 n02951585 can_opener 0.10530355
## 2 n03255030 dumbbell 0.10238248
## 3 n02786058 Band_Aid 0.10219256
## 4 n03804744 nail 0.07120511
## 5 n03481172 hammer 0.04246261Zdjęcie szympansa.
2 modele nie słusznie klasyfikują bardziej jako goryla.
plot_img('10.jpeg')
what_is_it('10.jpeg')## ResNet50
##
## class_name class_description score
## 1 n02480855 gorilla 0.790869117
## 2 n02481823 chimpanzee 0.075572461
## 3 n02492660 howler_monkey 0.065349638
## 4 n01688243 frilled_lizard 0.008603136
## 5 n02483708 siamang 0.008166263
##
##
## VGG16
##
## class_name class_description score
## 1 n02480855 gorilla 0.45563412
## 2 n02481823 chimpanzee 0.36622998
## 3 n02483708 siamang 0.05447859
## 4 n02133161 American_black_bear 0.03650714
## 5 n02492660 howler_monkey 0.03270785
##
##
## VGG19
##
## class_name class_description score
## 1 n02481823 chimpanzee 0.666312814
## 2 n02480855 gorilla 0.162429020
## 3 n02483708 siamang 0.131579682
## 4 n02493793 spider_monkey 0.021770868
## 5 n02492660 howler_monkey 0.008632749Zdjęcie Jana Pawła
O dziwo model rozpoznał że na zdjęciu jest rzeźba. Mamy wśród wyników piedestał, obelisk, megalit. Jeden rozpoznał materiał: miedź.
plot_img('11.jpg')
what_is_it('11.jpg')## ResNet50
##
## class_name class_description score
## 1 n03903868 pedestal 0.79254013
## 2 n02892201 brass 0.07897896
## 3 n04486054 triumphal_arch 0.04248532
## 4 n03837869 obelisk 0.02953168
## 5 n02747177 ashcan 0.01614135
##
##
## VGG16
##
## class_name class_description score
## 1 n03903868 pedestal 0.42331406
## 2 n03837869 obelisk 0.21745521
## 3 n04346328 stupa 0.06972346
## 4 n03743016 megalith 0.05373014
## 5 n04486054 triumphal_arch 0.02554311
##
##
## VGG19
##
## class_name class_description score
## 1 n03837869 obelisk 0.27158356
## 2 n03743016 megalith 0.15461192
## 3 n03903868 pedestal 0.14220390
## 4 n04486054 triumphal_arch 0.04620933
## 5 n04346328 stupa 0.03491951Snikers
Tu ku mojemu zaskoczeniu 3 model bardzo jest pewny że na zdjęciu znajduje się harmonijka.
plot_img('13.jpg')
what_is_it('13.jpg')## ResNet50
##
## class_name class_description score
## 1 n03871628 packet 0.25860724
## 2 n03908618 pencil_box 0.16058408
## 3 n06794110 street_sign 0.11610079
## 4 n04548362 wallet 0.04848527
## 5 n02840245 binder 0.04121584
##
##
## VGG16
##
## class_name class_description score
## 1 n03494278 harmonica 0.39240646
## 2 n03871628 packet 0.20743132
## 3 n04548362 wallet 0.10290398
## 4 n06794110 street_sign 0.06403392
## 5 n07248320 book_jacket 0.03200327
##
##
## VGG19
##
## class_name class_description score
## 1 n03494278 harmonica 0.83302629
## 2 n03908618 pencil_box 0.03172375
## 3 n04548362 wallet 0.02809962
## 4 n02877765 bottlecap 0.02366120
## 5 n03871628 packet 0.02038583